- The paper introduces an algorithm that needs no knowledge of prediction error and achieves additive loss O(min(√b·max(EMD,1), b log b)), combining O(√b) consistency with O(b log b) robustness.
- The method delays the predicted optimal purchase time and truncates overly long rentals using predicted-tail structure, while a λ-interpolated variant offers a tunable balance between prediction-dependent performance and classical competitiveness.
- The paper proves matching lower bounds showing that linear EMD dependence is impossible, the √b·max(EMD,1) scaling is tight within its natural form, and O(√b) consistency necessarily entails Ω(b log b) robustness.
Problem setting and motivation
This paper studies the classical ski rental problem within the algorithms-with-predictions framework, replacing the standard point prediction with a distributional prediction. The setup is as follows: rental costs 1 per day, buying costs b>1, and the true number of ski days T is drawn from an unknown distribution p over [N]. The algorithm receives a predicted distribution p^, measured against p via Earth Mover's Distance (EMD, i.e., Wasserstein-1). A deterministic policy AK rents for K days and buys on day K+1; the benchmark is the expected cost of the optimal policy for the true distribution, T0, and additive loss is T1.
Two framing choices matter. First, benchmarking against T2 rather than the optimum in hindsight is necessary: for a bimodal truth placing mass T3 at T4 and T5 at T6, any policy incurs T7 loss relative to hindsight even when T8 is known exactly. Second, collapsing T9 to a point (via sampling or MLE) can be arbitrarily lossy for multimodal predictions, motivating a genuinely distributional treatment.
The key distinction from prior work concerns what the algorithm knows about prediction quality. Besbes, Ma, and Mouchtaki assume p0 is known to the algorithm; this enables stronger bounds of p1 but makes robustness trivial (one can always fall back to buying immediately for worst-case loss p2). This paper removes that assumption — answering an open question posed explicitly by Besbes et al. — which forces robustness to become a genuine design constraint.
Main upper bound
The central result is an algorithm achieving, without any knowledge of p3,
p4
This yields consistency p5 (additive loss at zero error) and robustness p6 (additive loss at unbounded error), simultaneously and without knowing p7. Notably, whenever p8 the loss is strictly sublinear in p9, beating the trivial bound.
The algorithm has two components:
- Base algorithm: compute the optimal policy [N]0 for [N]1, then delay purchase by [N]2 additional days. The analysis compares [N]3 to [N]4 through the optimal transport plan between [N]5 and [N]6, classifying transport maps as "good" (self-bounding by their [N]7 contribution) or "bad" (split into short-distance moves, controlled by structural properties of optimal policies in [N]8, and long-distance moves, whose total mass is at most [N]9).
- Truncation step for robustness: let p^0 be the first day where the tail mass of p^1 beyond p^2 is at most p^3; if p^4, buy at day p^5 instead. The truncation loss analysis uses a clean tail lemma: for any p^6 and p^7, p^8, so the probability that truncation triggers under the truth is at most p^9, bounding extra cost by p0.
Robustness rests on two structural lemmas about the predicted tails p1: optimality of p2 forces the predicted tail to decay geometrically before p3 (with ratio p4 per step backwards), and dyadic tail levels must be crossed within windows of length p5. Together these imply the truncation time is always p6 regardless of how misleading p7 is. The authors note the base-algorithm bound also follows from equation (E-28) of Besbes et al.'s appendix, though their proof was developed independently and is arguably more direct.
Further robustification
For applications demanding worst-case guarantees below p8, the paper gives a p9-interpolated algorithm, following Purohit et al.: it clamps the base algorithm's buying time toward day EMD0 from either side, buying at day EMD1 or EMD2 as appropriate. It satisfies
EMD3
where EMD4. At EMD5 this recovers the main theorem; at EMD6 it recovers the classical factor-EMD7 competitive ratio. This mirrors the point-prediction tradeoff of Purohit et al., generalized to distributions.
Matching lower bounds
The paper supplies four lower bounds, all holding against randomized algorithms (via Yao's lemma), establishing near-optimality of both the bound's shape and the consistency–robustness tradeoff.
| Result |
Statement |
| No linear-in-EMD generalization |
No algorithm achieves loss EMD8 for any EMD9 |
| Necessity of AK0 |
No algorithm achieves loss AK1 |
| Tightness of AK2 |
No algorithm achieves loss AK3 |
| Pareto-optimal tradeoff |
Consistency AK4 forces robustness AK5 |
Several points deserve emphasis. First, the impossibility of AK6 means the point-prediction guarantee of Purohit et al. does not extend to distributions: error and AK7 must combine nonlinearly. Second, the third lower bound matches the main upper bound exactly within the natural family AK8, so the non-robust component of the bound cannot be improved in this form. Third, the Pareto-optimality result is proved via an explicit adversarial prediction whose tail decays at rate AK9 up to K0 (making K1 optimal) and more slowly thereafter; any algorithm consistent on this prediction must rent until K2, suffering K3 loss when the truth is a point mass at K4.
There are caveats. The tightness of K5 holds only within the restricted functional form K6; the paper does not rule out better bounds of a different shape. Also, in the limited setting where the truth is a point distribution, the linear-in-EMD lower bound does not apply: the paper gives a deterministic algorithm achieving loss K7, derandomizing the sampling-based approach.
Relative to the three closest papers, the contribution is well delineated. Purohit, Svitkina, and Kumar supply the point-prediction template (including the K8-interpolation technique reused here). Diakonikolas et al. address sample complexity of learning the true distribution within structured families and provide no error-dependent guarantees. Besbes et al. study essentially the same problem under known K9 via robust optimization; their stronger K+10 bound exploits knowledge of K+11, and their explicit open question — performance without knowledge of K+12 — is precisely what this paper resolves. The work also connects to binary search with distributional predictions, which posed a similar question in a different problem.
Limitations and open questions
The paper concedes several boundaries. The robustness term K+13 carries an extra logarithmic factor over the trivial worst-case loss of K+14 achievable when K+15 is known; whether unknown-quality robustness below K+16 is possible while retaining sublinear consistency remains unresolved except through the K+17-interpolation, which degrades the low-error regime. The matching lower bound for K+18 is conditional on the multiplicative form of the bound, so a qualitatively different (e.g., additive-plus-multiplicative) improvement is not excluded. Finally, the analysis assumes deterministic policies for the upper bounds and finite support K+19, and the benchmark choice excludes comparisons to the optimum in hindsight, where T00 loss is unavoidable.
Conclusion
This paper settles the ski rental problem with distributional predictions of unknown quality: an T01-consistent, T02-robust algorithm exists without knowledge of the prediction error, and its consistency–robustness tradeoff is Pareto-optimal, with the loss function itself essentially tight among natural forms. The results close an open question of Besbes et al. and establish that the transition from point to distributional predictions genuinely changes the achievable landscape — linear dependence on prediction error is impossible, and the correct scaling is T03 capped at T04.